A novel layered data reduction mechanism for clustering fMRI data

Abstract

Original fMRI data often contains a variety of noise caused by the operator, the equipment, the environment, etc. To suppress the noise, many processing methods based on smoothing have been proposed to analyze the fMRI data. In this study, a layered data reduction mechanism is presented to alleviate the influence of noise while retaining the spatial construction of the fMRI data. The layered data reduction method consists of two layers criteria to reduce the noise voxels: (a) the isolated voxel; (b) the isolated generated cube. The 1-layer data reduction procedure aims to remove all those isolated voxels whose corresponding generated cube only contains one single voxel under a preset threshold ξ. The 2-layer data reduction procedure is aimed at removing those isolated generated cubes whose corresponding final cube only contains one single generated cube. A simplified genetic algorithm (SGA) is proposed to determine the optimum threshold ξ adaptively. Meanwhile, to avoid that some useful information would be lost on account of all the isolated voxels and isolated generated cubes being reduced directly, a compensation mechanism taking advantage of multivariate RV measure is used to decrease the probability of incorrect data reduction. The classical FCM (Fuzzy c-means) method is adopted to cluster the data having been implemented by the layered data reduction method. Extensively experimental results show that the proposed layered data reduction method is effective and can efficiently improve the clustering accuracy on the hybrid data and the real fMRI data.